广州励美生活垃圾AI分类数据库
收藏资源简介：
通过矩阵相机对通过分类管道里的物质进行不间断拍照，采集后的图片经物理存贮转至实验室。首先筛选出具有回收及再加工价值的物质如矿泉水瓶、塑料饭盒等，再对其进行标注及定义，为后期机器学习和目标检测提供支持。 生活垃圾中的物质有其特殊性，物质会变形、污染或掺杂其它易附物质如泥土等等，因此，单一类物质如矿泉水瓶需通过增加学习大量的图像特征来提升其分类或检测能力。首先，矿泉水瓶的样本必须足够大，除管道日采集以外，还需通过人为干预，对漏检的在实验室里通过人为手段变换矿泉水瓶的角度、外形等等再采图标注作补充。其次，利用基于卷积神经网络和cuDNN的算法，通过数据增强、HSV空间变化、亮度调整、镜像调整、裁剪等手段，使用来训练深度学习模型文件的图片数据集更高效准确。目前模型的准确率达7成以上，换言之，视觉系统利用此数据库及其生成的模型，使机器人和PLC对在人工智能分选带上的十种物质抓取率达70%以上。
Continuous photography of materials passing through the classification pipeline is performed using a matrix camera. The captured images are physically stored and transferred to the laboratory. First, materials with recycling and reprocessing value such as mineral water bottles and plastic lunch boxes are screened out, then annotated and defined, to provide support for subsequent machine learning and object detection tasks. Materials in domestic waste have unique characteristics: they may be deformed, contaminated, or mixed with other easily adhered substances such as soil. Therefore, for single-type materials like mineral water bottles, their classification or detection capabilities need to be improved by learning a large volume of image features. First, the sample size of mineral water bottles must be sufficiently large. Apart from the collection via the pipeline on a daily basis, manual intervention is also required: for the missed-detected samples, their angles, shapes and other attributes are manually adjusted in the laboratory for re-photography and annotation to supplement the dataset. Second, algorithms based on convolutional neural networks and cuDNN are employed, with techniques including data augmentation, HSV space transformation, brightness adjustment, mirroring, cropping and other methods, to make the image datasets used for training deep learning models more efficient and accurate. Currently, the accuracy of the model exceeds 70%. In other words, the vision system leveraging this database and the generated models enables robots and PLCs to achieve a grasping rate of over 70% for the ten types of materials on the AI sorting conveyor belt.
